A new energy time-sharing multi-day power generation capacity prediction system and method based on a large model
By using a large model-based approach, new energy power generation areas are divided and climate feature maps are extracted. These maps are then fused with time trend features to solve the problem of dynamic changes in climate features in new energy power generation forecasting, thus achieving efficient power generation capacity forecasting.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI LVSI SHU INNOVATION ENERGY TECHNOLOGY CO LTD
- Filing Date
- 2025-08-07
- Publication Date
- 2026-05-15
AI Technical Summary
Existing new energy power generation prediction models are unable to accurately capture the dynamic changes in natural climate characteristics and lack an efficient feature fusion mechanism, resulting in prediction results that cannot fully reflect the actual power generation situation.
A large model-based approach is adopted to divide the new energy power generation area into rectangular sub-regions. Climate feature maps are extracted through convolutional neural networks, and dimensionality reduction and attention matrix updates are performed to obtain spatial correlation characteristics. These characteristics are then fused with temporal trend features, and prediction is made using a Transformer encoder and a multilayer perceptron.
It has achieved accurate prediction of new energy power generation capacity, improved the effective utilization rate of spatial features and the ability to identify time series sequence dependencies, captured multi-scale periodic features, and improved the accuracy of prediction.
Smart Images
Figure CN120952244B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent power generation management technology, specifically a new energy time-sharing multi-day power generation capacity prediction system and method based on a large model. Background Technology
[0002] With the rapid development of new energy sources, new energy power generation forecasting plays a crucial role in power system dispatch and energy management. Currently, new energy power generation forecasting typically relies on time-series analysis of historical power generation data to uncover trends and predict future generation capacity. However, in different regions, natural climate characteristics can interfere with new energy power generation, making it difficult for traditional forecasting methods to accurately capture these dynamic changes.
[0003] On the other hand, existing prediction models often lack efficient feature fusion mechanisms when dealing with multi-source heterogeneous data, such as meteorological data and power generation data, resulting in prediction results that cannot fully reflect the actual power generation situation. Summary of the Invention
[0004] The purpose of this invention is to provide a new energy time-of-use multi-day power generation capacity prediction system and method based on a large model, so as to solve the problems raised in the prior art.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the time-of-use and multi-day power generation capacity of new energy sources based on a large model, the method comprising:
[0006] Step S100: The power supply area of a certain new energy power generation is recorded as the target area. The target area is divided into several rectangular sub-regions. Through the geographical adjacency relationship and climate correlation between the rectangular sub-regions, a climate feature map with sub-regions as nodes is obtained.
[0007] Step S200: Extract hierarchical spatiotemporal features of the climate feature map based on convolutional neural network, perform dimensionality reduction on the climate data in the target area, update the weight matrix to a dynamic attention matrix through normalization operation, obtain the spatial correlation characteristics between different sub-regions, and obtain the spatial feature matrix of the target area.
[0008] Step S300: Embed time location codes into the historical sequence of power generation to obtain the time trend features of power generation, and fuse the time trend features with spatial correlation characteristics to obtain the spatiotemporal fusion features of the target area;
[0009] Step S400: The spatiotemporal location coding is fused with the spatiotemporal features. Multi-scale periodic features are captured through the encoder hierarchy. Then, the predicted features are extracted from the encoder output and the predicted features are mapped to the predicted power generation value using a multilayer perceptron.
[0010] Furthermore, step S100 includes:
[0011] Step S101: Divide the target area into N rectangular sub-regions according to the equal-spacing grid rule, and collect all the rectangular sub-regions into the sub-region set R, R={r1, r2, r3, ..., rN}, where r1, r2, r3, ... and rN represent the 1st, 2nd, 3rd, ... and Nth rectangular sub-regions, respectively;
[0012] Step S102: Obtain the i-th rectangular sub-region ri and the j-th rectangular sub-region rj in the sub-region set R, and set the adjacent elements A of ri and rj. ij When ri and rj have a common boundary, A ij =1, when ri and rj do not have a common boundary, A ij =0, collect all the adjacency elements between any two rectangular subregions to obtain the adjacency matrix AC;
[0013] Step S103: Collect historical climate data and calculate the Pearson correlation coefficient c between the natural data of ri and rj. ij The Pearson correlation coefficients between any two rectangular sub-regions are aggregated to construct a correlation matrix C. The natural data include temperature, humidity, wind speed, sunshine duration, light intensity, water flow or water level difference.
[0014] Step S104: For a rectangular sub-region, correspond to a node in the graph structure, collect all nodes and record them in the node set VR, and calculate the edge weights s of the nodes corresponding to ri and rj. ij s ij =A ij ×c ij Collect all edge weights to obtain the edge set SE, and construct the climate feature graph G, G=(VR,SE).
[0015] A unified geographic unit division system is established to provide standardized spatial units for subsequent spatial correlation analysis. Then, the spatial topology of sub-regions is quantified to provide a geographic adjacency basis for climate feature maps. Furthermore, the spatiotemporal dependencies of climate indicators in different sub-regions are captured to provide climate correlation weights for climate feature maps. Geographic spatial relationships and climate correlation characteristics are integrated into structured map data to provide input for subsequent neural network extraction of spatiotemporal features.
[0016] Furthermore, step S200 includes:
[0017] Step S201: Perform parallel convolution operations on the climate feature map using convolution kernels of at least two sizes to obtain climate features at at least two spatial scales;
[0018] Step S202: The climate features at all spatial scales are stitched together to obtain the dimensions of the climate features. The dimensions of the climate features are compressed through a fully connected layer to obtain a dimension-reduced feature matrix.
[0019] Step S203: Based on the features of the rectangular sub-regions after dimensionality reduction, calculate and normalize the attention weights between the rectangular sub-regions, and collect the normalized attention weights to obtain the attention weight matrix;
[0020] Step S204: Multiply the attention weight matrix with the dimensionality reduction feature matrix to obtain the spatial feature matrix H of the fused spatial correlation.
[0021] Furthermore, step S300 includes:
[0022] Step S301: Set a unit time period, take the power generation in the target area within one unit time period as a power generation sample value, collect the power generation sample values of M consecutive unit time periods, and obtain the power generation sequence Y;
[0023] Step S302: Set the time encoding dimension dp and time position pos, where dp is a positive even number, pos is an integer, and satisfies 1≤pos≤M. Set the encoding dimension index b, 0≤b<dp / 2.
[0024] Step S303: Calculate the time position code of time position pos, where the even-numbered bits are encoded as P. pos,2b , ,
[0025] Odd-numbered bits are encoded as P pos,2b+1 , ,
[0026] The DIM table represents the time scale. ,
[0027] Traverse all time position codes from 1 to M for all unit periods, and collect all time position codes to obtain a time position code matrix P, where each row of the time position code matrix P corresponds to all time position codes for a time position.
[0028] Step S304: Expand the dimensions of the power generation sequence Y using linear projection to obtain the projection matrix Y of Y. proj Y proj =Y·WP T Where WP represents the projection matrix, and the time trend feature matrix T is calculated, T=Y proj +P;
[0029] To integrate the power generation value Y and the periodic information P of time location, the dimensions of Y and P need to be aligned. P has a size of M×dp, and the power generation samples are arranged into an M×1 power generation sequence Y. Therefore, Y needs to be projected into a projection matrix Y of size M×dp using a projection matrix WP of size dp×1. proj ;
[0030] Step S305: Collect the elements of the i-th row in the spatial feature matrix H to form the feature vector h of the i-th region. i Collect all the region feature vectors, and perform average pooling on all the region feature vectors in the feature matrix H to obtain the pooled vector S. avg ;
[0031] Step S306: Calculate the pooling vector S avg The projection matrix St, St=S avg ·E, where E is a single-row vector of all 1s with M elements. The spatiotemporal splicing matrix X is obtained by concatenating the time trend feature matrix T and the projection matrix St. st .
[0032] By using global average pooling or attention mechanisms, spatial dimensional features are aggregated into a single spatial context vector, reducing spatial dimensional complexity. The spatial context vector is then extended along the time dimension to match the time trend feature dimension. Finally, the time trend features are concatenated with the extended spatial features to form a feature matrix that integrates spatiotemporal information, providing a foundation for subsequent predictions.
[0033] Furthermore, step S400 includes:
[0034] Step S401: Perform a linear transformation on the spatiotemporal encoding matrix P to obtain the projection matrix P of P. proj P proj =P·W d +b d W d Let b represent the linear transformation matrix. d Indicates the bias term;
[0035] Step S402: Overlay the temporal encoding with the spatial and spatiotemporal fusion features to obtain the sequence spatiotemporal location information X. pos X pos =X st +P proj ;
[0036] Step S403: Extract the output of the last layer of the Transformer encoder as the prediction feature F containing spatiotemporal and periodic information. pred F pred =Encoder(X pos), where Encoder represents the encoder in the Transformer architecture;
[0037] Step S404: Predict the features F pred Input MLP, mapped to the predicted power generation value y, y=W o ·ReLU (W h ·F pred +b h )+b o W o and W h Represents the MLP weights, ReLU represents the ReLU activation function, and b h and b o This indicates the bias term.
[0038] Location encoding injects temporal location information into spatiotemporal features, facilitating model recognition of time sequence. The Transformer encoder extracts multi-scale dependencies and nonlinear features from the time series through a multi-head self-attention mechanism and a feedforward neural network. Through the encoder's hierarchical structure, it captures periodic features at different time scales layer by layer. Feature vectors containing multi-period information are extracted from the encoder output as the basis for prediction. The multilayer perceptron maps the predicted features to specific predicted power generation values, completing the final prediction task.
[0039] To better implement the above methods, a new energy time-sharing multi-day power generation capacity prediction system based on a large model is also proposed. The system includes: a climate feature map management module, a spatial feature management module, a spatiotemporal fusion feature management module, and a power generation prediction module.
[0040] The climate feature map management module is used to manage the climate feature map of the target area. The spatial feature management module is used to capture the spatial correlation characteristics between different sub-regions and obtain the spatial feature matrix of the target area. The spatiotemporal fusion feature management module is used to fuse the temporal trend characteristics with the spatial correlation characteristics to obtain the spatiotemporal fusion characteristics of the target area. The power generation prediction module is used to predict power generation through encoders and multilayer perceptrons.
[0041] Furthermore, the climate feature map management module includes: sub-region management unit, adjacency management unit, natural data management unit, and node feature management unit;
[0042] The sub-region management unit manages rectangular sub-regions of the target area, the adjacency management unit manages the adjacency relationships of the rectangular sub-regions, the natural data management unit manages the natural data of the target area, and the node feature management unit aggregates node features and edge weights to construct a climate feature map.
[0043] Furthermore, the spatial feature management module includes: a spatial scale extraction unit, a feature compression unit, an attention weight management unit, and a spatial association fusion unit;
[0044] The spatial scale extraction unit is used to extract the climate features of the target area at the spatial scale through convolution kernels. The feature compression unit is used to compress the dimension of the climate features through a fully connected layer to obtain a dimension-reduced feature matrix. The attention weight management unit is used to calculate and normalize the attention weights between rectangular sub-regions, and to collect the normalized attention weights to obtain the attention weight matrix. The spatial association fusion unit is used to fuse spatial associations to obtain a spatial feature matrix.
[0045] Furthermore, the spatiotemporal fusion feature management module includes: a power generation data management unit, a time coding management unit, a time coding embedding unit, and a spatiotemporal feature fusion unit;
[0046] Among them, the power generation data management unit is used to manage the power generation of the target area, the time coding management unit is used to manage the time location coding of the power generation sequence, the time coding embedding unit is used to embed time coding into the power generation sequence, and the spatiotemporal feature fusion unit is used to obtain a spatiotemporal splicing matrix by combining time trend features and regional features.
[0047] Furthermore, the power generation prediction module includes: a sequence spatiotemporal feature management unit, a prediction feature management unit, and a power generation prediction value management unit;
[0048] The sequence spatiotemporal feature management unit is used to overlay time coding with spatial and spatiotemporal fusion features to obtain sequence spatiotemporal location information; the prediction feature management unit is used to extract prediction features through the encoder in the Transformer architecture; and the power generation prediction value management unit is used to calculate the power generation prediction value.
[0049] Compared with the prior art, the beneficial effects of the present invention are:
[0050] By standardizing natural data and then integrating it, a high-quality data base is provided. A mechanism of "spatial global aggregation + temporal dimension alignment" is introduced to dynamically match cross-regional spatial features with temporal features through temporal dimension expansion.
[0051] Spatial features are aggregated using attention mechanisms or global average pooling, and high-influence regions are automatically focused through attention weights, thereby improving the effective utilization rate of spatial features.
[0052] By combining positional encoding technology, the model can accurately identify the sequential dependencies of time series. It adopts a Transformer multi-layer encoder architecture to achieve joint capture of features from short-period to long-period changes, thereby improving the feature capture capability of single temporal change properties. Attached Figure Description
[0053] Figure 1 This is a schematic diagram of the structure of a new energy time-of-use multi-day power generation capacity prediction system based on a large model according to the present invention;
[0054] Figure 2 This is a flowchart illustrating a method for predicting the time-of-use, multi-day power generation capacity of new energy sources based on a large model, according to the present invention. Detailed Implementation
[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0056] Example: Figures 1-2 As shown, the present invention provides a technical solution: a new energy time-sharing multi-day power generation capacity prediction system and method based on a large model;
[0057] The methods include:
[0058] Step S100: The power supply area of a certain new energy power generation is recorded as the target area. The target area is divided into several rectangular sub-regions. Through the geographical adjacency relationship and climate correlation between the rectangular sub-regions, a climate feature map with sub-regions as nodes is obtained.
[0059] Step S100 includes:
[0060] Step S101: Divide the target area into N rectangular sub-regions according to the equal-spacing grid rule, and collect all the rectangular sub-regions into the sub-region set R, R={r1, r2, r3, ..., rN}, where r1, r2, r3, ... and rN represent the 1st, 2nd, 3rd, ... and Nth rectangular sub-regions, respectively;
[0061] Step S102: Obtain the i-th rectangular sub-region ri and the j-th rectangular sub-region rj in the sub-region set R, and set the adjacent elements A of ri and rj. ij When ri and rj have a common boundary, A ij =1, when ri and rj do not have a common boundary, A ij =0, collect all the adjacency elements between any two rectangular sub-regions to obtain the adjacency matrix AC;
[0062] Step S103: Collect historical climate data and calculate the Pearson correlation coefficient c between the natural data of ri and rj. ijThe Pearson correlation coefficients between any two rectangular sub-regions are aggregated to construct a correlation matrix C. The natural data include temperature, humidity, wind speed, sunshine duration, light intensity, water flow or water level difference.
[0063] Step S104: For a rectangular sub-region, correspond to a node in the graph structure, collect all nodes and record them in the node set VR, and calculate the edge weights s of the nodes corresponding to ri and rj. ij s ij =A ij ×c ij Collect all edge weights to obtain the edge set SE, and construct the climate feature graph G, G=(VR,SE).
[0064] In this embodiment, natural data is used, and T sampling times are set to sample the indices of the natural data of the i-th rectangular sub-region ri and the j-th rectangular sub-region rj. The index of the natural data of the rectangular sub-region ri at the t-th sampling time is denoted as x. it The index of the natural data for the rectangular subregion rj is denoted as x. jt ;
[0065] Calculate the Pearson correlation coefficient c between the i-th rectangular subregion ri and the j-th rectangular subregion rj. ij ,
[0066] The Pearson correlation coefficients between any two rectangular subregions are collected to obtain the correlation coefficient matrix C.
[0067] Step S200: Extract hierarchical spatiotemporal features of the climate feature map based on convolutional neural network, perform dimensionality reduction on the climate data in the target area, update the weight matrix to a dynamic attention matrix through normalization operation, obtain the spatial correlation characteristics between different sub-regions, and obtain the spatial feature matrix of the target area.
[0068] Step S200 includes:
[0069] Step S201: Perform parallel convolution operations on the climate feature map using convolution kernels of at least two sizes to obtain climate features at at least two spatial scales;
[0070] Step S202: The climate features at all spatial scales are stitched together to obtain the dimensions of the climate features. The dimensions of the climate features are compressed through a fully connected layer to obtain a dimension-reduced feature matrix.
[0071] Step S203: Based on the features of the rectangular sub-regions after dimensionality reduction, calculate and normalize the attention weights between the rectangular sub-regions, and collect the normalized attention weights to obtain the attention weight matrix;
[0072] Step S204: Multiply the attention weight matrix with the dimensionality reduction feature matrix to obtain the spatial feature matrix H of the fused spatial correlation.
[0073] The metrics of natural data collected from all rectangular sub-regions are used to construct the node input feature matrix X;
[0074] A K-layer convolutional network is constructed, where K=3 in the embodiment, and the kernel size of each convolutional layer is 3×3, 5×5 and 7×7 respectively.
[0075] Parallel extraction is performed on the node input feature matrix X. , where Wk represents the convolution kernel of the k-th convolutional layer, bk is the bias term of the k-th convolutional layer, and σ is the activation function. In the embodiments, k takes the values 1, 2, and 3 respectively.
[0076] Calculate the dimensionality reduction feature matrix Z, Z = Wz·concat(F1, F2, F3) + bz, where concat represents the concatenation function, and Wz and bz are the parameters of the dimensionality reduction layer;
[0077] Obtain the dimensionality-reduced features of the i-th rectangular sub-region ri and the j-th rectangular sub-region rj, denoted as zi and zj respectively;
[0078] Calculate the attention weights eij for the i-th and j-th rectangular sub-regions.
[0079] W a and U a Let represent the weight matrices of the i-th and j-th rectangular sub-regions, respectively, and let a be the attention vector;
[0080] Collect the attention weights of the N-1 rectangular sub-regions (excluding the i-th rectangular sub-region) and the i-th rectangular sub-region, and normalize the attention weights of the i-th and j-th rectangular sub-regions to obtain the normalized attention weight α. ij ,
[0081] The attention weight matrix AN is obtained by collecting the normalized attention weights.
[0082] Calculate the feature matrix H of the fused spatial correlation, H = AN·Z.
[0083] Step S300: Embed time location codes into the historical sequence of power generation to obtain the time trend features of power generation, and fuse the time trend features with spatial correlation characteristics to obtain the spatiotemporal fusion features of the target area;
[0084] Step S300 includes:
[0085] Step S301: Set a unit time period, take the power generation in the target area within one unit time period as a power generation sample value, collect the power generation sample values of M consecutive unit time periods, and obtain the power generation sequence Y;
[0086] Step S302: Set the time encoding dimension dp and time position pos, where dp is a positive even number, pos is an integer, and satisfies 1≤pos≤M. Set the encoding dimension index b, 0≤b<dp / 2.
[0087] Step S303: Calculate the time position code of time position pos, where the even-numbered bits are encoded as P. pos,2b , ,
[0088] Odd-numbered bits are encoded as P pos,2b+1 , ,
[0089] The DIM table represents the time scale. ,
[0090] Traverse all time position codes from 1 to M for all unit periods, and collect all time position codes to obtain a time position code matrix P, where each row of the time position code matrix P corresponds to all time position codes for a time position.
[0091] Step S304: Expand the dimensions of the power generation sequence Y using linear projection to obtain the projection matrix Y of Y. proj Y proj =Y·WP T Where WP represents the projection matrix, and the time trend feature matrix T is calculated, T=Y proj +P;
[0092] Step S305: Collect the elements of the i-th row in the spatial feature matrix H to form the feature vector h of the i-th region. i Collect all the region feature vectors, and perform average pooling on all the region feature vectors in the feature matrix H to obtain the pooled vector S. avg ;
[0093] Step S306: Calculate the pooling vector S avg The projection matrix St, St=S avg ·E, where E is a single-row vector of all 1s with M elements. The spatiotemporal splicing matrix X is obtained by concatenating the time trend feature matrix T and the projection matrix St. st .
[0094] Step S400: Fuse the spatiotemporal location code with the spatiotemporal features, capture multi-scale periodic features through the encoder hierarchy, extract the predicted features from the encoder output, and use a multilayer perceptron to map the predicted features to the predicted power generation value.
[0095] Step S400 includes:
[0096] Step S401: Perform a linear transformation on the spatiotemporal encoding matrix P to obtain the projection matrix P of P. proj P proj =P·W d +b d W d Let b represent the linear transformation matrix. d Indicates the bias term;
[0097] Step S402: Overlay the temporal encoding with the spatial and spatiotemporal fusion features to obtain the sequence spatiotemporal location information X. pos X pos =X st +P proj ;
[0098] Step S403: Extract the output of the last layer of the Transformer encoder as the prediction feature F containing spatiotemporal and periodic information. pred F pred =Encoder(X pos ), where Encoder represents the encoder in the Transformer architecture;
[0099] Step S404: Predict the features F pred Input MLP, mapped to the predicted power generation value y, y=W o ·ReLU (W h ·F pred +b h )+b o W o and W h Represents the MLP weights, ReLU represents the ReLU activation function, and b h and b o This indicates the bias term.
[0100] The query vector Qc, key vector Kc, and value vector Vc of the c-th head attention are calculated using linear projection, where Qc is the query vector Qc, the key vector Kc, and the value vector Vc. c =X pos ·W qc K c =X pos ·W kc V c =X pos ·W vc Wqc W kc and W vc These represent the query vector Q. c Key vector K c Sum vector V c The projection weight matrix;
[0101] Calculate the attention of the c-th head. c , Where dk = (dp + d) / C, d is the dimension of the natural data, and C is the total number of attention heads;
[0102] Furthermore, feature representation capabilities are enhanced through feedforward neural networks and layer normalization, and the training process is stabilized to compute the input X in the Transformer architecture. out X out =LayerNorm(X pos +MultiHead(X pos ))+FFN(LayerNorm(X pos +MultiHead(X pos (), where LayerNorm represents layer normalization, FFN represents feedforward neural network, MultiHead(X) pos ) represents multi-head attention splicing, MultiHead(X) pos ) = concat(Head1, Head2, Head3,…, Head C );
[0103] X is processed through the L-layer encoder of the Transformer. out Perform hierarchical enhancement, and denote the output of the Lth layer as X. L out ;
[0104] , ,
[0105] Fpred=λFlast+(1-λ)Favg, λ∈(0,1).
[0106] The system includes:
[0107] The module includes a climate feature map management module, a spatial feature management module, a spatiotemporal fusion feature management module, and a power generation prediction module.
[0108] The climate feature map management module is used to manage the climate feature map of the target area. The spatiotemporal fusion feature management module includes: a power generation data management unit, a time coding management unit, a time coding embedding unit, and a spatiotemporal feature fusion unit.
[0109] Among them, the power generation data management unit is used to manage the power generation of the target area, the time coding management unit is used to manage the time location coding of the power generation sequence, the time coding embedding unit is used to embed time coding into the power generation sequence, and the spatiotemporal feature fusion unit is used to obtain a spatiotemporal splicing matrix by combining time trend features and regional features.
[0110] The spatial feature management module is used to capture the spatial correlation characteristics between different sub-regions and obtain the spatial feature matrix of the target region. The spatial feature management module includes: a spatial scale extraction unit, a feature compression unit, an attention weight management unit, and a spatial correlation fusion unit.
[0111] The spatial scale extraction unit is used to extract the climate features of the target area at the spatial scale through convolution kernels. The feature compression unit is used to compress the dimension of the climate features through a fully connected layer to obtain a dimension-reduced feature matrix. The attention weight management unit is used to calculate and normalize the attention weights between rectangular sub-regions, and to collect the normalized attention weights to obtain the attention weight matrix. The spatial association fusion unit is used to fuse spatial associations to obtain a spatial feature matrix.
[0112] The spatiotemporal fusion feature management module is used to fuse time trend features with spatial correlation characteristics to obtain the spatiotemporal fusion features of the target area. The spatiotemporal fusion feature management module includes: a power generation data management unit, a time coding management unit, a time coding embedding unit, and a spatiotemporal feature fusion unit.
[0113] Among them, the power generation data management unit is used to manage the power generation of the target area, the time coding management unit is used to manage the time location coding of the power generation sequence, the time coding embedding unit is used to embed time coding into the power generation sequence, and the spatiotemporal feature fusion unit is used to obtain a spatiotemporal splicing matrix by combining time trend features and regional features.
[0114] The power generation prediction module is used to predict power generation through an encoder and a multilayer perceptron. The power generation prediction module includes: a sequence spatiotemporal feature management unit, a prediction feature management unit, and a power generation prediction value management unit.
[0115] The sequence spatiotemporal feature management unit is used to overlay time coding with spatial and spatiotemporal fusion features to obtain sequence spatiotemporal location information; the prediction feature management unit is used to extract prediction features through the encoder in the Transformer architecture; and the power generation prediction value management unit is used to calculate the power generation prediction value.
[0116] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A method for predicting the time-of-use and multi-day power generation capacity of new energy sources based on a large model, characterized in that: The methods include: Step S100: The power supply area of a certain new energy power generation is recorded as the target area. The target area is divided into several rectangular sub-regions. Through the geographical adjacency relationship and climate correlation between the rectangular sub-regions, a climate feature map with sub-regions as nodes is obtained. Step S200: Extract hierarchical spatiotemporal features of the climate feature map based on the convolutional neural network, perform dimensionality reduction on the climate data in the target area, update the weight matrix to a dynamic attention matrix through normalization, obtain the spatial correlation characteristics between different sub-regions, and obtain the spatial feature matrix of the target area. Step S200 includes: Step S201: Perform parallel convolution operations on the climate feature map using convolution kernels of at least two sizes to obtain climate features at at least two spatial scales; Step S202: The climate features at all spatial scales are stitched together to obtain the dimensions of the climate features. The dimensions of the climate features are compressed through a fully connected layer to obtain a dimension-reduced feature matrix. Step S203: Based on the features of the rectangular sub-regions after dimensionality reduction, calculate and normalize the attention weights between the rectangular sub-regions, and collect the normalized attention weights to obtain the attention weight matrix; Step S204: Multiply the attention weight matrix with the dimensionality-reduced feature matrix to obtain the spatial feature matrix H that fuses spatial correlations; Step S300: Embed time location codes into the historical sequence of power generation to obtain the time trend features of power generation, and fuse the time trend features with spatial correlation characteristics to obtain the spatiotemporal fusion features of the target area; Step S300 includes: Step S301: Set a unit time period, take the power generation in the target area within one unit time period as a power generation sample value, collect the power generation sample values of M consecutive unit time periods, and obtain the power generation sequence Y; Step S302: Set the time encoding dimension dp and time position pos, where dp is a positive even number, pos is an integer, and 1≤pos≤M is satisfied. Set the encoding dimension index b, 0≤b<dp / 2; Step S303: Calculate the time position code of time position pos, where the even-numbered bits are encoded as P. pos,2b , , Odd-numbered bits are encoded as P pos,2b+1 , , The DIM table represents the time scale. ; Traverse all time position codes from 1 to M for all unit periods, and collect all time position codes to obtain a time position code matrix P, where each row of the time position code matrix P corresponds to all time position codes for a time position. Step S304: Expand the dimensions of the power generation sequence Y using linear projection to obtain the projection matrix Y of Y. proj Y proj =Y·WP T Where WP represents the projection matrix, and the time trend feature matrix T is calculated, T=Y proj +P; Step S305: Collect the elements of the i-th row in the spatial feature matrix H to form the feature vector h of the i-th region. i Collect all the region feature vectors, and perform average pooling on all the region feature vectors in the feature matrix H to obtain the pooled vector S. avg ; Step S306: Calculate the pooling vector S avg The projection matrix St, St=S avg ·E, where E is a single-row vector of all 1s with M elements. The spatiotemporal splicing matrix X is obtained by concatenating the time trend feature matrix T and the projection matrix St. st ; Step S400: The spatiotemporal location coding is fused with the spatiotemporal features. Multi-scale periodic features are captured through the encoder hierarchy. Then, the predicted features are extracted from the encoder output and the predicted features are mapped to the predicted power generation value using a multilayer perceptron.
2. The method for predicting the time-of-use, multi-day power generation capacity of new energy sources based on a large model, as described in claim 1, is characterized in that: Step S100 includes: Step S101: Divide the target area into N rectangular sub-regions according to the equal-spacing grid rule, and collect all the rectangular sub-regions into the sub-region set R, R={r1, r2, r3, ..., rN}, where r1, r2, r3, ... and rN represent the 1st, 2nd, 3rd, ... and Nth rectangular sub-regions, respectively; Step S102: Obtain the i-th rectangular sub-region ri and the j-th rectangular sub-region rj in the sub-region set R, and set the adjacent elements A of ri and rj. ij When ri and rj have a common boundary, A ij =1, when ri and rj do not have a common boundary, A ij =0, collect all the adjacency elements between any two rectangular sub-regions to obtain the adjacency matrix AC; Step S103: Collect historical climate data and calculate the Pearson correlation coefficient c between the natural data of ri and rj. ij The Pearson correlation coefficients between any two rectangular sub-regions are aggregated to construct a correlation matrix C. The natural data include temperature, humidity, wind speed, sunshine duration, light intensity, water flow or water level difference. Step S104: For a rectangular sub-region, correspond to a node in the graph structure, collect all nodes and record them in the node set VR, and calculate the edge weights s of the nodes corresponding to ri and rj. ij s ij =A ij ×c ij Collect all edge weights to obtain the edge set SE, and construct the climate feature graph G, G=(VR,SE).
3. The method for predicting the time-of-use, multi-day power generation capacity of new energy sources based on a large model, as described in claim 2, is characterized in that: Step S400 includes: Step S401: Perform a linear transformation on the spatiotemporal encoding matrix P to obtain the projection matrix P of P. proj P proj =P·W d +b d W d Let b represent the linear transformation matrix. d Indicates the bias term; Step S402: Overlay the temporal encoding with the spatial and spatiotemporal fusion features to obtain the sequence spatiotemporal location information X. pos X pos =X st +P proj ; Step S403: Extract the output of the last layer of the Transformer encoder as the prediction feature F containing spatiotemporal and periodic information. pred F pred =Encoder(X pos ), where Encoder represents the encoder in the Transformer architecture; Step S404: Predict the features F pred Input MLP, mapped to the predicted power generation value y, y=W o ·ReLU (W h ·F pred +b h )+b o W o and W h Represents the MLP weights, ReLU represents the ReLU activation function, and b h and b o This indicates the bias term.
4. A new energy time-of-use multi-day power generation capacity prediction system based on a large model, used to execute the new energy time-of-use multi-day power generation capacity prediction method based on a large model as described in any one of claims 1-3, characterized in that: The system includes: Climate feature map management module, spatial feature management module, spatiotemporal fusion feature management module, and power generation prediction module; The climate feature map management module is used to manage the climate feature map of the target area. The spatial feature management module is used to capture the spatial correlation characteristics between different sub-regions and obtain the spatial feature matrix of the target area. The spatiotemporal fusion feature management module is used to fuse the temporal trend characteristics with the spatial correlation characteristics to obtain the spatiotemporal fusion characteristics of the target area. The power generation prediction module is used to predict power generation through encoders and multilayer perceptrons.
5. A new energy time-of-use multi-day power generation capacity prediction system based on a large model according to claim 4, characterized in that: The climate feature map management module includes: sub-region management unit, adjacency management unit, natural data management unit, and node feature management unit; The sub-region management unit manages rectangular sub-regions of the target area, the adjacency management unit manages the adjacency relationships of the rectangular sub-regions, the natural data management unit manages the natural data of the target area, and the node feature management unit aggregates node features and edge weights to construct a climate feature map.
6. The new energy time-sharing multi-day power generation capacity prediction system based on a large model according to claim 4, characterized in that: The spatial feature management module includes: a spatial scale extraction unit, a feature compression unit, an attention weight management unit, and a spatial correlation fusion unit; The spatial scale extraction unit is used to extract the climate features of the target area at the spatial scale through convolution kernels. The feature compression unit is used to compress the dimension of the climate features through a fully connected layer to obtain a dimension-reduced feature matrix. The attention weight management unit is used to calculate and normalize the attention weights between rectangular sub-regions, and to collect the normalized attention weights to obtain the attention weight matrix. The spatial association fusion unit is used to fuse spatial associations to obtain a spatial feature matrix.
7. A new energy time-of-use multi-day power generation capacity prediction system based on a large model according to claim 4, characterized in that: The spatiotemporal fusion feature management module includes: a power generation data management unit, a time coding management unit, a time coding embedding unit, and a spatiotemporal feature fusion unit; Among them, the power generation data management unit is used to manage the power generation of the target area, the time coding management unit is used to manage the time location coding of the power generation sequence, the time coding embedding unit is used to embed time coding into the power generation sequence, and the spatiotemporal feature fusion unit is used to obtain a spatiotemporal splicing matrix by combining time trend features and regional features.
8. A new energy time-of-use multi-day power generation capacity prediction system based on a large model according to claim 4, characterized in that: The power generation prediction module includes: a sequence spatiotemporal feature management unit, a prediction feature management unit, and a power generation prediction value management unit; The sequence spatiotemporal feature management unit is used to overlay time coding with spatial and spatiotemporal fusion features to obtain sequence spatiotemporal location information; the prediction feature management unit is used to extract prediction features through the encoder in the Transformer architecture; and the power generation prediction value management unit is used to calculate the power generation prediction value.